What it means
Pretraining is the first and most compute-intensive stage of building a large AI model: the model is trained on huge volumes of text, code, images, or other data to learn general patterns and skills before it is specialized for any task. It works by repeatedly predicting missing or next tokens across trillions of examples, adjusting billions of internal weights each step. This is where raw compute capacity, clusters of GPUs or accelerators running for weeks or months, is converted into base-model capability, so pretraining sits at the top of the AI workload stack and drives demand for chips, high-bandwidth memory, networking, and power. It is a lever because better data, scale, and efficiency yield more capable models, and a constraint because it is bottlenecked by available compute, memory bandwidth, and interconnect between the accelerators.
Why it matters to investors
Pretraining is what turns capital spent on compute into frontier-model capability, so its scale sets the demand for accelerators, memory, and networking beneath it. The labs that pretrain competitive base models, including DeepSeek, Moonshot AI, and Safe Superintelligence, anchor that demand, though most remain private.
Companies on this part of the chain
Named to show where the term sits in the AI supply chain — research, not advice, and never a recommendation to buy or sell.
Related terms
See Pretraining in the live AI chain.
THE ENTITY maps every constraint onto one live model — which part is tight now, who owns it, and who gets squeezed when it moves. Plain-English reads you can check.
THE ENTITY is an educational read on the AI supply chain — research, not investment advice. It explains how the chain works and who sits where, never price targets or buy/sell calls.